The CMU Machine Translation Systems at WMT 2013: Syntax, Synthetic Translation Options, and Pseudo-References

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 70–77, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics The CMU Machine Translation Systems at WMT 2013: Syntax, Synthetic Translation Options, and Pseudo-References Waleed Ammar Victor Chahuneau Michael Denkowski Greg Hanneman Wang Ling Austin Matthews Kenton Murray Nicola Segall Yulia Tsvetkov Alon Lavie Chris Dyer∗ Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA ∗Corresponding author: cdyer@cs.cmu.edu Abstract We describe the CMU systems submit- ted to the 2013 WMT shared task in ma- chine translation. We participated in three language pairs, French–English, Russian– English, and English–Russian. Our particular innovations include: a label- coarsening scheme for syntactic tree-to- tree translation and the use of specialized modules to create “synthetic translation options” that can both generalize beyond what is directly observed in the parallel training data and use rich source language context to decide how a phrase should translate in context. 1 Introduction The MT research group at Carnegie Mellon Uni- versity’s Language Technologies Institute par- ticipated in three language pairs for the 2013 Workshop on Machine Translation shared trans- lation task: French–English, Russian–English, and English–Russian. Our French–English sys- tem (§3) showcased our group’s syntactic sys- tem with coarsened nonterminal types (Hanne- man and Lavie, 2011). Our Russian–English and English–Russian system demonstrate a new multi- phase approach to translation that our group is us- ing, in which synthetic translation options (§4) to supplement the default translation rule inven- tory that is extracted from word-aligned training data. In the Russian-English system (§5), we used a CRF-based transliterator (Ammar et al., 2012) to propose transliteration candidates for out-of- vocabulary words, and used a language model to insert or remove common function words in phrases according to an n-gram English language model probability. In the English–Russian system (§6), we used a conditional logit model to predict the most likely inflectional morphology of Rus- sian lemmas, conditioning on rich source syntac- tic features (§6.1). In addition to being able to generate inflected forms that were otherwise unob- served in the parallel training data, the translations options generated in this matter had features re- flecting their appropriateness given much broader source language context than usually would have been incorporated in current statistical MT sys- tems. For our Russian–English system, we addition- ally used a secondary “pseudo-reference” transla- tion when tuning the parameters of our Russian– English system. This was created by automatically translating the Spanish translation of the provided development data into English. While the output of an MT system is not always perfectly gram- matical, previous work has shown that secondary machine-generated references improve translation quality when only a single human reference is available when BLEU is used as an optimization criterion (Madnani, 2010; Dyer et al., 2011). 2 Common System Components The decoder infrastructure we used was cdec (Dyer et al., 2010). Only the constrained data resources provided for the shared task were used for training both the translation and language models. Word alignments were generated us- ing the Model 2 variant …

Content truncated to 500 words